S4PRED — ONNX weights for in-browser inference
ONNX export of the S4PRED secondary structure predictor by Moffat & Jones, for use with ONNX Runtime Web.
Files
| File | Size | Description |
|---|---|---|
s4pred_int8.onnx | 225 MB | int8 weight-quantised ensemble (recommended) |
s4pred_fp16.onnx | 449 MB | fp16 ensemble |
Both files contain the full 5-model GRU ensemble. The int8 variant quantises GRU weight matrices to int8 via direct initialiser surgery, retaining float32 activations — prediction accuracy is effectively identical to fp32 (0.3% of residues differ across 100 test proteins).
Live app
Try it in your browser (no install, no data sent to server): zacharyardern.github.io/s4pred-wasm
Original model
Moffat, L. & Jones, D.T. (2021). Increasing the accuracy of single sequence prediction methods using a deep semi-supervised learning framework. Bioinformatics, 37(21), 3744–3750. https://doi.org/10.1093/bioinformatics/btab491
Source code and original weights: https://github.com/psipred/s4pred
Original weights are © the authors, released under GPL-3.0. This repo redistributes them in ONNX format under the same licence.
Export method
# Float32 export
torch.onnx.export(model, dummy, 's4pred.onnx', dynamo=False, opset_version=17,
dynamic_axes={'tokens': {1: 'seq_len'}, 'log_probs': {0: 'seq_len'}})
# int8: direct GRU initialiser quantisation (quantize_dynamic skips GRU nodes)
# see quantize_gru_weights.py in https://github.com/ZacharyArdern/s4pred-wasm